akkiisfrommars
@akkiisfrommars
Followers
794
Following
20K
Media
358
Statuses
14K
19 | Building foundational LLMs @MistyozAI
Hyd
Joined August 2020
We’re launching CosmicFish - India’s first foundational LLM built fully from scratch. Models are now live on Hugging Face Likes and reposts are appreciated ♥️
9
15
79
Will be postponed to November! Would be better to launch it along side the android app
0
0
0
Just got a crazy idea for a new hybrid Transformer-SSM model!! Super excited to build it.
0
0
1
CosmicChat early October updates: - Smarter CosmicFish models - Markdown Rendering - Faster Inference - Internet lookup beta
1
0
3
Gonna make my own mini search engine and implement it into CosmicChat so that CosmicFish 300M can get the latest info from the internet.
0
0
4
iOS 26 seems to have made CosmicChat’s inference improve. I’m able to notice a difference in the generation speed of CosmicFish 300M
0
0
2
Damn CosmicFish 300M is good at context based instruction following. Should get even better though…
1
1
4
What if we had agents that use custom CosmicFish variants like 300M, 120M and 90M to do a particular task? Like basically use all three or more models simultaneously all running locally on your phone.
0
0
2
Mixture Of Experts would work so good on mobile. Definitely have to try making CosmicFish-MOE work well
0
0
3
Bruhhhhh why is Apple taking so long to approve my app update Like I can’t wait to announce CosmicChat
0
0
2
Could a 200M parameter MOE LLM work? Something like 4x50M
1
0
4
If everything goes right, I’ll probably launch CosmicChat in 2 days 🤞
2
0
7
Found some interesting tek India's 1st foundational LLM = CosmicFish Built entirely from scratch by @akkiisfrommars, this is set to revolutionise how LLM's are trained and interact. Best part? It can be used offline! What's an LLM, you ask? Peep the 2nd frame (You're welcome)
4
3
19
Hmmm maybe I should train a new CosmicFish model where admitting that it doesn’t know something is also rewarded.
OpenAI realesed new paper. "Why language models hallucinate" Simple ans - LLMs hallucinate because training and evaluation reward guessing instead of admitting uncertainty. The paper puts this on a statistical footing with simple, test-like incentives that reward confident
0
0
9
Just started working on CosmicFish 1.5 Gonna be some pretty major architectural changes and will significantly improve the performance Working towards a CosmicFish 1.5 90M model in the coming weeks… Will be added to CosmicChat :)
1
1
6